A heterogeneous memory system is a memory system where several different memories, or levels of memory, are used to satisfy memory demands of an computer application. An example of an application for a heterogeneous memory system is in graphics processing systems. Different levels of memory are used by a graphics processing system to facilitate graphics processing and rendering of graphics images on a display. A first level of memory is typically embedded memory that is fabricated directly on the same semiconductor substrate as a graphics processor. Embedded memory can provide data to the graphics processor at very low access times, and consequently, increase the speed at which graphics data may be processed. A second level of memory is typically memory that is external to the device, but located on the same graphics card as the graphics processor. Memory such as this is commonly referred to as external, or local memory. A third level of memory is AGP memory, or host memory that the graphics processor can access through a system bus. Host memory generally has the greatest access time of the three levels of memories because the graphics processor can only access the AGP memory via a system bus and several different memory and bus controllers. Although local memory can provide data more quickly than the host memory, it still is considerably slower than the embedded memory of the first level of memory.
For a conventional heterogeneous memory system, there are two typical arrangements. A first example of a heterogeneous memory system is arranged with a single memory controller to handle all memory accesses. Such an arrangement is illustrated in FIG. 1. The memory system 10 includes a central memory controller 12 coupled to both memory 20 through memory bus 16, and memory 22 through memory bus 18. The memory 20 may be representative of embedded memory, and the memory 22 may be representative of external memory. In operation, the central memory controller 12 receives memory access requests from various requesting entities, such as a graphics processor, over buses 14a–n. The central memory controller 12 services the various memory access requests by determining whether the requested memory address is located in the memory 20 or the memory 22. The appropriate memory device is accessed and data is written to or read therefrom. An arrangement such as memory system 10 has the advantage that additional memory may be easily added because all memory access requests are serviced by the central memory controller 12. For the same reason, the various memory access requests can all be handled seamlessly by the central memory controller 12. That is, when a memory access request is made, only the central memory controller 12 must determine which memory, either memory 20 or memory 22, to access. However, a problem with the arrangement of memory system 10 is that there are physical limitations as to the number of buses 14a–n that may be routed to the memory controller 12. Additionally, as the complexity of the central memory controller 12 increases to accommodate a greater number of memory access requests, the amount of space the central memory controller occupies also increases. Thus, space overhead issues become a concern in applications where small graphics processing systems are desired.
A second example of a heterogeneous memory system is shown in FIG. 2. Memory system 30 addresses some of the concerns raised by the memory system 10 of FIG. 1. The memory system 30 includes a central memory controller 12 coupled to a memory 20 through a memory bus 16. The central memory controller 12 services only the memory access requests made to memory 20. The memory system 30 also includes memory 22 directly coupled to a requesting entity through memory bus 32. Thus, memory access requests to memory 22 may be only made over the memory bus 32. The memory 20 may represent embedded memory, while the memory 22 may represent local memory. As illustrated in FIG. 2, all memory access requests to memory 20 are controlled by the central memory controller 12. However, access to the memory 22, is controlled directly by the requesting entity coupled to the bus 32. That is, access to memory 22 can be made only by the requesting entity hardwired to the bus 32.
The memory system 30 does, to some degree, resolve the issues with regards to the physical limitations of routing a plurality of request lines to a single central memory controller, as well as space overhead issues resulting from the complexity of using a central memory controller. However, a problem with the memory system 30 is that the allocation of available memory is fixed according to the design of the circuitry. That is, the memory 22 may be accessed only by the requesting entity to which it is coupled through bus 32. Any available memory in the memory 22 cannot be reallocated for another purpose, such as storing overflow data from the memory 20. Furthermore, memory access requests must be delegated prior to being made either to the central memory controller 12 or the memory 22, rather than having all memory access requests simply handled by a single central memory controller. Moreover, adding additional memory to the memory system 30 is made more difficult by the fixed arrangement. Additional memory cannot simply be reallocated, but must be added to supplement either memory 20 or memory 22, but not both.
Therefore, there is a need for a memory system where the number of memory access request lines to a memory controller is reduced and where the available memory may be allocated efficiently.